Your proof package
Your model result is only believable when someone can see how you tested it. Put the evaluation in the repo and in the demo description, not only in your head.
The minimum proof package is: dataset, split method, baseline, confusion matrix, key metric, and one limitation.
Where it goes
Use two places. First, write an Evaluation section in your README or project description. Second, mention the same result in your demo script or video.
If judges are short on time, they should still understand what the model was tested on and what it cannot do yet.
Add the section
Do each one yourself, then tap it to tick it off. The ticks are only a checklist for you: they are not marked or scored.
0 of 5 done
README template
Copy this structure. Replace the sample values with your own numbers. Do not keep placeholder text.
Model Evaluation
Dataset:
Used 2,400 labelled support messages collected from our demo dataset. Labels: refund, delivery, product, other.
Split method:
80/20 random split with a fixed seed. Train: 1,920 rows. Validation: 480 rows.
Baseline:
Majority-class baseline predicts "delivery" for every message. Baseline accuracy: 42%.
Model result:
Validation accuracy: 78%.
Key metric: macro F1 = 0.74, because classes are not equally common.
Confusion matrix:
Rows are actual labels. Columns are predicted labels.
refund delivery product other
refund 82 11 5 6
delivery 14 151 13 10
product 8 21 72 9
other 6 19 4 49
Limitation:
The model was tested only on English messages. It may perform worse on Hinglish or regional language messages.Do not oversell
Do not write “our model is 95% accurate” without the split, baseline, and confusion matrix. That number can be useless if the data is imbalanced or leaked.
Do not hide weak classes. If the model misses many refund complaints or fraud cases, say so and explain the risk.
Use AI carefully
You can ask ChatGPT, Gemini, or Claude to turn your raw results into a short evaluation paragraph. Paste only the numbers you already computed.
Check that the AI does not invent a dataset size, metric, or limitation. If a number is not in your notebook or script output, do not include it.
Demo script line
Say the result in plain words during the demo. This sounds more credible than only showing a high score.
We tested the model on a held-out 20% validation split. It beats the majority-class baseline, 78% accuracy versus 42%. The confusion matrix shows it still confuses product and delivery messages, so we would not use it for automatic replies without human review.Do this now
Do each one yourself, then tap it to tick it off. The ticks are only a checklist for you: they are not marked or scored.
0 of 5 done